SwinDSR: Thangka Super-Resolution Model Based on DCT Domain Feature Enhancement and DCT based HQ Image Reconstruction

Thangka, a form of traditional Tibetan painting, holds deep cultural value and showcases a unique artistic essence. Deep learning-based image super-resolution technology is an effective means to restore low-resolution Thangka images. This technology not only supports the digital preservation of Thangka images but also promotes the global spreading and study of Thangka culture, further enriching the depth of related art and historical research. Nevertheless, current image super-resolution techniques are inadequate for Thangka images due to the following reason: Compared to other publicly available datasets, Thangka images are rich in content, with highly complex textures and colors, making it challenging for current methods to produce high-quality reconstruction, especially in terms of high-frequency details. To overcome these challenges, a Image Super-Resolution Using Swin Transformer And Discrete Cosine Transform (SwinDSR) is proposed, which has a larger receptive field and reconstructs high-definition images in the DCT domain. The method includes two parts: (1) a Discrete Cosine Transform Domain Feature Enhancement Module (DFEM) to expand the receptive field, extract global features, and enhance image quality and structure; (2) a Discrete Cosine Transform based HQ Image Reconstruction (DHR) to avoid the reliance on local pixels during high-resolution image reconstruction, fully utilizing global information while enhancing high-frequency details to improve the consistency of details between the reconstructed Thangka image and the original image. Qualitative and quantitative experiments confirm the proposed method achieves excellent super-resolution effect on Thangka images. The proposed method can contribute to the restoration of low-resolution Thangka images.